Hybrid control for combining model-based and model-free reinforcement learning

We develop an approach to improve the learning capabilities of robotic systems by combining learned predictive models with experience-based state-action policy mappings. Predictive provide understanding task and dynamics, while (model-free) mappings encode favorable actions that override planned actions. refer our systematically model-based model-free methods as hybrid learning. Our efficiently learns motor skills improves performance policies. Moreover, enables policies (both model-free) be updated using any off-policy reinforcement method. derive a deterministic method optimally switching between modalities. adapt stochastic variation relaxes some key assumptions in original derivation. variations are tested on variety robot control benchmark tasks simulation well hardware manipulation task. extend for use imitation methods, where experience is provided through demonstrations, we test expanded capability real-world pick-and-place The results show capable improving sample efficiency experimental domains.